You hand the same documents over to AI agents that evaluate the output.
The same rules that you, an expert would apply, should be documented and used by an agent to apply those heuristics.
There are obviously some areas where there's spot judgement, deviations, and exceptions. But by and large, you have some heuristic that you use when evaluating output. Document it and give it to an agent.
This suffices for merely random error (as long as the error is demonstrable) but it's not a general solution to the problem.